【作者】 杨振宇；

【导师】 李伟；

【作者基本信息】 中国农业大学 ， 机械制造及其自动化， 2014， 博士

【摘要】 从设施农业自动化的角度出发,针对钵苗在移栽过程中,需要对其品质进行人工分选的现状。以穴盘钵苗为研究对象,钵苗的移栽适合度和苗叶调向角度为研究目标,研究了基于机器视觉的钵苗移栽适合度信息获取方法和钵苗苗叶调向方法,设计了一种钵苗自动分选移栽机软硬件系统,并对其关键结构部件进行了优化设计。本论文的主要研究内容和研究成果如下：(1)研究了基于图像颜色空间的钵苗图像分割方法。通过增加单色背景板和顶升钵苗的方式获取视频图像,分别对四种颜色空间的钵苗试验样本在不同光照条件下的图像分割稳定性进行研究。将钵苗HSI颜色空间中S分量提取与最大类间方差阈值分割算法相结合的方法用于钵苗视频图像分割,减少了图像处理的难度,提高了钵苗图像分割的效率和鲁棒性,可满足钵苗图像茎叶提取对算法实时性和稳定性的要求。(2)研究了移栽钵苗的高度信息获取方法。在对钵苗分割后图像获取钵苗高度信息的三种识别算法分析的基础上,通过综合评价三种钵苗高度识别算法的平均相对偏差和算法运行效率,采用实时性和鲁棒性更好的顶点链码主轴法最小外接矩形算法获取钵苗高度。(3)研究了移栽钵苗的直立度信息获取方法。在对分割图像进行骨架提取、水平1×5膨胀和垂直13x1腐蚀提取主茎秆特征后,分析了根据主茎杆特征获取钵苗直立度信息的三种识别算法。通过综合评价三种钵苗直立度识别算法的平均相对偏差和算法运行效率,采用实时性和鲁棒性更好的基于Susan角点直线拟合算法获取钵苗直立度。(4)研究了钵苗苗叶调向方法。以钵苗绕顶杆转轴旋转1圈的48帧视频图像为目标对象,提出了基于核函数兴趣点相似性度量跟踪的苗叶调向方法。利用分割、骨架提取后图像的中心像素8邻域判别获取钵苗苗叶兴趣点,通过平滑相似度函数和目标候选最小化函数有效确定当前帧兴趣点的位置,采用最小化接近一致性代价幻影点函数产生幻影点代替可能丢失的兴趣点完成钵苗兴趣点的跟踪,依据兴趣点在图像平面内的x坐标值和变化规律可实现钵苗苗叶方向的有效调整。(5)提出了基于变分光流的钵苗视频帧图像关键帧提取方法。为了减少兴趣点跟踪时处理视频帧图像的数量,将钵苗视频帧图像的光流求解转化为能量泛函求极值,采用基于Laplacian守恒假设的数据项和基于各向同性光流驱动的平滑项,通过比较各相邻帧图像间光流矢量的马氏距离均值可有效从48帧视频图像中获取关键帧图像序列。(6)设计了钵苗自动分选移栽机软硬件系统,优化了其关键部件。该机由底座支架、平带输送单元、钵苗顶升旋转单元、钵苗移栽单元、视频图像信息获取单元和系统控制与图像处理单元等部件组成。设计了钵苗自动分选移栽机系统软件,软件可在连续动态工作模式下运行,可实现对不同种类钵苗的信息获取与判别。提出了一种基于模糊优化算法的钵苗顶升旋转机构优化设计方法,使被约束的物理量可处在具有模糊边界的范围内,与常用优化算法相比更符合工程实际。更多还原

【Abstract】 Viewed from the automation of facility agriculture production, to solve the problem of require artificial quality separation in potted-seedlings transplanting process, this dissertation takes potted-seedling as the research object, potted-seedling transplanting fitness and angle modulation of potted-seedlings leaves as the research goals, studied the methods of potted-seedlings transplanting fitness information acquisition and potted-seedlings leaves adjustment direction using machine vision, designed hardware and software systems of potted-seedling automatic sorting transplanter, and optimized design of it’s key component. The main research contents and conclusions were as follows:(1) Method of potted-seedling image segmentation was studied based on color space. The stability of image segmentation was study respectively to the four color space for potted-seedling test samples with different light conditions by increasing the monochrome background panels and lifting of potted-seedling obtained video images. Threshold segmentation algorithm was determined that combining the S component extraction of HSI color space and OTSU for video image segmentation of potted seedlings. It could reduce the difficulty of image processing, and improve the efficiency of potted-seedling image segmentation, recognition and robustness, and meet real-time performance and stability requirements that the stems and leaves of potted-seedling image extraction algorithm.(2) Method of potted-seedling height information acquisition was studied. Three recognition algorithm of potted-seedling height information was analysed using the potted-seedling segmentation images. Method of potted-seedling height acquisition was determined that minimum external rectangle algorithm based on vertex chain code spindle after evaluate synthetically relative deviation and algorithm efficiency of three kinds of potted-seedling height recognition algorithm.(3) Method of potted-seedling perpendicularity information acquisition was studied. Three recognition algorithm of potted-seedling perpendicularity information was analysed using the main stem of potted-seedling was extracted after thinning,1×5horizontal dilation and13×1vertical erosion image. Method of potted-seedling perpendicularity acquisition was determined that straight line fitting method of corner based on SUSAN corner detection algorithm after evaluate synthetically relative deviation and algorithm efficiency of three kinds of potted-seedling perpendicularity recognition algorithm.(4) Method of potted-seedling leaves adjustment direction was studied. Taking48frame video image of rotating potted-seedling as the research object, potted-seedling leaves adjustment direction method was proposed using kernel function interest points similarity measure trace. Interest points of potted-seedling leaves was obtained using image center pixel8neighborhood identifying after segmentation and thinning. Interest points location of the current frame video image was effectively determined by smoothing similarity function and target candidate minimized function. Missing interest points were replaced by minimized costs closer to consistency function produces the phantom points for following the tracks of potted-seedling interest points. According to x coordinate and variation of interest points in the image plane, and the direction of potted-seedling leaves could be availably adjusted.(5) Method of key frame video images extraction was presented using variation optical flow, in order to reduce the number of video images when tracing interest points. Optical flow computation of potted-seedling video images was converted into the energy functional extremum using Laplacian conserved hypothetical data and isotropic optical flow driven smoothly. Key frame sequence can be effectively obtained from48-frame video images by comparing mahalanobis distance values of the optical flow vector between two adjacent images.(6) Hardware system and software system of potted-seedling automatic sorting transplanter were designed, and key components were optimized. The set consists of mounting bracket,8flat belt conveyor, potted-seedlings lift&rotate unit, transplanting information acquisition unit and system control unit, video image processing unit and other components. Optimization design method of key components was proposed for potted-seedling lifting rotation mechanism based on fuzzy optimal algorithm, physical quantities were limited to the range with the blurred boundaries, and compared with the commonly optimization algorithms more practical. Software system was designed, and it can be run in continuous mode. It can realize acquisition and judgment of different kinds of potted-seedlings transplanting information.更多还原